JP4381455B2 - Character size estimation method, apparatus, and recording medium - Google Patents

Description

  The present invention relates to a character size estimation method and apparatus for estimating a character size in an image, and a recording medium on which a character size estimation processing program is recorded.

  When character recognition or the like is performed, the character size is extracted as preprocessing. For example, a character size is extracted using projection data in each region obtained by dividing a document image into strips (see Patent Document 1), and the vertical width and horizontal width of a pixel block are determined from the peripheral distribution of the document image. A character size is extracted by calculation (see Patent Document 2), a long character, a true character, and a plain character are discriminated, and a character size is determined based on the width / height of the character (see Patent Document 3) Refer to), when estimating the character size in the first direction of the handwritten character string, from the second direction size after the first merge, select the nth dimension value from the larger one, and based on this, the character size Examples include those for obtaining an estimated value (see Patent Document 4) and those for estimating a character size from an average value of white run lengths (see Patent Document 5).

Japanese Patent No. 2569151 JP-A-5-89283 JP-A-5-282492 Japanese Patent Laid-Open No. 7-21312 JP-A-7-184034

  By the way, conventionally, when a table is processed, the character size included in the table is not estimated in advance, and if it is less than the planned character size, it is less likely to be erroneously recognized as a line. However, in a table including a character size larger than the planned character size, there is a high possibility that a straight line component existing in the character is erroneously recognized as a ruled line.

  An object of the present invention is to provide a character size estimation method, apparatus, and recording medium for accurately estimating the character size in an image.

In order to achieve the above object, according to the first aspect of the present invention, the step of extracting the run in the main scanning direction and the sub-scanning direction of the input image, the step of obtaining the frequency distribution of the extracted run, A step of setting a predetermined threshold based on a frequency distribution of runs, a step of extracting a connected rectangle using a run larger than the predetermined threshold, a step of extracting a ruled line from the extracted connected rectangle, and a predetermined region A step of determining a character rectangle when a predetermined number of ruled lines are present in the main scanning direction and the sub-scanning direction, a step of obtaining a frequency distribution of a vertical size of the rectangle determined to be the character rectangle, and detecting a peak of the frequency distribution of the vertical size, when searching the direction of larger vertical size from the peak, with the step of outputting the vertical size of the frequency values is equal to or less than a predetermined value as a character size It is characterized by a door.

According to the second aspect of the present invention, a unit for extracting a run in the main scanning direction and the sub-scanning direction of an input image, a unit for obtaining a frequency distribution of the extracted run, and a predetermined value based on the frequency distribution of the run Means for setting a threshold value, means for extracting a connected rectangle using a run larger than the predetermined threshold value, means for extracting a ruled line from the extracted connected rectangle, a main scanning direction and a sub-scan in a predetermined area When there are more than a predetermined number of ruled lines in the direction, a means for determining a character rectangle, a means for obtaining a frequency distribution of the vertical size of the rectangle determined to be the character rectangle, and a peak of the frequency distribution of the vertical size It is characterized by comprising means for detecting and means for outputting, as a character size, a vertical size with a frequency value equal to or less than a predetermined value when searching in a direction in which the vertical size is larger from the peak .

The invention described in claim 3 is a computer-readable recording medium storing a program for causing a computer to implement the character size estimation method described in claim 1 .

According to the present invention , since a linear component existing in a character is used, it is possible to determine a character rectangle that is resistant to noise. In addition, it is possible to estimate while absorbing the variation in the character size of the character rectangle that exists most in the image.

According to the present invention , it is possible to extract a short ruled line that cannot be extracted with a conventional fixed threshold. Further, since the distribution itself does not depend on the size of the table, a large table or a small table can be processed by the same processing method. That is, the processing method of the present invention does not depend on the resolution.

According to the present invention , since the threshold is estimated from the differential value of the distribution, more accurate threshold estimation is possible.

According to the present invention , since the high-frequency noise in the distribution is removed, a more accurate threshold value can be estimated.

According to the present invention, it is possible to estimate an appropriate threshold value not only for a table frame including vertical lines and horizontal lines, but also for a distribution consisting of simple lines.

According to the present invention , since attribute information is given to a run, the original image erasure process of the run corresponding to the character can be performed at high speed.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

Example 1
FIG. 1 shows a configuration of the first embodiment of the present invention, and FIG. 2 shows a processing flowchart of the first embodiment. In the figure, 1 is an image input unit, 2 is an original picture memory, 3 is a run extraction unit, 4 is a connected rectangle extraction unit, 5 is a frequency counting unit, 6 is a peak detection unit, and 7 is a character size output unit.

  Hereinafter, the processing operation of the first embodiment will be described with reference to FIG. The document is read by the image input unit 1 such as a scanner, and the input image is stored in the original image memory 2 (step 101). The run extraction unit 3 extracts a run in the main scanning direction (or sub-scanning direction) of the image data in the original image memory 2 and stores it in the memory (step 102).

  Next, the connected rectangle extraction unit 4 extracts connected rectangles using the extracted runs in the main scanning direction (step 103). The frequency counting unit 5 counts the frequency of the extracted rectangular vertical size (or horizontal size) (step 104). The peak detector 6 searches for a point where the sign of the differential value changes from the smaller vertical size on the frequency distribution, and sets this point as a peak (step 105). The character size output unit 7 outputs the above peak as the character size (step 106).

  Thus, by using the peak of the frequency distribution, it is possible to estimate the size of the most characters in the image.

  In the above-described embodiment, the character size is estimated at the peak of the vertical and horizontal distribution of the rectangle, but certain character sizes are not all the same size, but vary depending on the characters. Therefore, in order to absorb this variation, the character size is estimated by the end value of the vertical and horizontal distribution of the rectangle. That is, after the peak is detected, the larger vertical size from the peak is searched, and the point at which the frequency falls below a certain value is set as the character size.

  Furthermore, when using a plurality of character sizes, in order to estimate the plurality of character sizes, after searching for a peak, for all the searched peaks, a search is made for the larger vertical size from the peak, and the frequency The point where is below a certain value is the character size.

  FIG. 3 shows an example of the vertical size frequency distribution of a character rectangle including two character sizes. Regarding the characters of the same character size, the horizontal size of the extracted connected rectangles varies widely, but the vertical size has a characteristic that falls within a certain range as shown in the figure. This characteristic does not depend on kanji or English. And the character size which exists in the read image is estimated by finding the area | region (two area | regions in a figure) which is a lump of distribution.

  That is, in the example of FIG. 3, when the peak is output as the character size, 40 (dots) is estimated as the character size. Further, when searching for the larger vertical size from the peak and setting the character size to a point where the frequency is less than a certain value, in the example of FIG. 3, the point where the frequency is less than a certain value (for example, 2), That is, 45 (dot) is estimated as the character size. Furthermore, when a plurality of character sizes are estimated, 65 (dots), which is a point at which the frequency becomes a certain value (for example, 2) or less, is also estimated as the character size.

(Example 2)
In the second embodiment, connected rectangle extraction is performed prior to table processing and the like. In this embodiment, it is determined in advance whether or not the extracted rectangle is a character, thereby estimating the character size more accurately. is there. In addition, the character size is estimated more accurately by determining the character rectangle using the linear component included in the character.

  FIG. 4 shows a configuration of the second embodiment of the present invention, and FIG. 5 shows a processing flowchart of the second embodiment. In FIG. 4, 21 is an image input unit, 22 is an original image memory, 23 is a run extracting unit, 24 is a connected rectangle extracting unit, 25 is a ruled line extracting unit, 26 is a character rectangle determining unit, 27 is a frequency counting unit, and 28 is a peak. A detection unit and a 29 character size output unit.

  The processing operation of the second embodiment will be described below with reference to FIG. The document is read by the image input unit 21 such as a scanner, and the input image is stored in the original image memory 22 (step 201). The run extraction unit 23 extracts a run in the main scanning direction of the image data in the original image memory 22 and stores it in the memory (step 202).

  Next, the connected rectangle extracting unit 24 extracts connected rectangles only for runs that are larger than a predetermined threshold (fixed threshold) with respect to the runs extracted in the main scanning direction (step 203). A ruled line (straight line component) is extracted from the extracted connected rectangle (step 204). Similar processing is performed in the sub-scanning direction (step 206), and ruled lines are extracted.

  If there are three or more ruled lines in both the main scanning direction and the sub-scanning direction (step 207), the character rectangle determination unit 26 determines that the character rectangle is a character rectangle (step 208). The above processing is performed for all rectangles (step 209).

  The frequency counting unit 27 counts the frequency for the vertical size of the rectangle determined as the character rectangle (step 210). The peak detector 28 searches for a point where the sign of the differential value changes from the smaller vertical size on the frequency distribution, and sets this point as a peak (step 211). The character size output unit 29 searches for the larger vertical size from the above-mentioned peak, and outputs the point where the frequency is below a certain value as the character size (step 212).

(Example 3)
In this embodiment, even if the character rectangles are in contact with each other and a rectangle having a size exceeding the estimated character size is formed, the character rectangle is estimated based on the number of ruled lines per unit area. That is, assuming that the area of the rectangular area having the estimated character size as one unit is one unit, the number of ruled lines per unit area is calculated for a connected rectangle larger than the rectangular area, and the character rectangle is determined from the number of ruled lines. .

  FIG. 6 shows the configuration of the third embodiment of the present invention, and FIGS. 7 and 8 show processing flowcharts of the third embodiment. In the third embodiment, a connected rectangle extracting unit 30, a ruled line extracting unit 31, and a character rectangle determining unit 32 are further added to the configuration of the second embodiment. In the processing flowchart of FIG. 8, the processing up to step 312 is the same as in the second embodiment. However, the character rectangle determined in step 308 is a temporary character rectangle.

  The following processing is performed for all rectangles determined to be temporary character rectangles. The connected rectangle extracting unit 30 extracts connected rectangles only for runs larger than the fixed threshold in the main scanning direction (step 313), and the ruled line extracting unit 31 extracts ruled lines (straight line components) from the extracted connected rectangles. Extract (step 314). Similar processing is performed in the sub-scanning direction to extract ruled lines.

  The character rectangle determination unit 32 divides the number of ruled lines in the main scanning direction / sub-scanning direction by (the rectangular area currently being processed / the area of the estimated character size), and the ruled lines per unit area (square of dots). The number is obtained (step 315), and if there are three or more ruled lines per unit area in either the main scanning direction or the sub-scanning direction, it is determined as a character rectangle (step 316).

(Example 4)
Example 4 is an example in which a more accurate character size is estimated by determining a character rectangle by core line processing. FIG. 9 shows the configuration of the fourth embodiment of the present invention, and FIG. 10 is a process flowchart of the fourth embodiment of the present invention. In the figure, 40 is an image input unit, 41 is an original image memory, 42 is a run extracting unit, 43 is a connected rectangle extracting unit, 44 is an ID assigning unit, 45 is a core rectangle extracting unit, 46 is a character rectangle determining unit, and 47 is a frequency. A counting unit, 48 is a peak detection unit, and 49 is a character size output unit.

  The document is read by the image input unit 40 such as a scanner, and the input image is stored in the original image memory 41 (step 401). The run extraction unit 42 extracts a run in the main scanning direction of the image data in the original image memory 41 and stores it in the memory (step 402). The connected rectangle extracting unit 43 extracts a connected rectangle using a run on the memory, and the ID assigning unit 44 assigns a rectangle ID (serial number) to the connected rectangle, and the rectangle ID constitutes the connected rectangle component. All the runs are also given (step 403).

  The core line rectangle extraction unit 45 extracts a rectangle for the runs having the same rectangle ID by using the core line of only the midpoint of the run (step 404), and performs the same processing in the sub-scanning direction to extract the core line rectangle. (Step 406). FIG. 11 shows an example of a core wire rectangle.

  If there are three or more core rectangles in both the main scanning direction and the sub-scanning direction (step 407), the character rectangle determining unit 46 determines that the character rectangle is a character rectangle (step 408). This process is performed for all rectangles (step 409). Thereafter, processing is performed in the same manner as in the second embodiment to output the character size.

(Example 5)
In the conventional method, ruled lines are extracted using a fixed threshold value. For this reason, it is difficult to extract a line having a length slightly larger than the size of the characters included in the table. This is because it is necessary to set a fixed threshold value that is large to some extent so that ruled lines are not extracted in characters in any document. In this way, in the conventional method, since a certain fixed threshold value is set to some extent, extraction of pseudo ruled lines in characters can be suppressed, but conversely, short ruled lines that are slightly larger than the character size are extracted. Can not do it.

  Therefore, in this embodiment, the threshold is not a fixed value but is estimated from the characteristics of the read document, and the ruled line is determined based on the threshold.

  FIG. 12 shows the configuration of the fifth embodiment. FIG. 13 is a process flowchart of the fifth embodiment. The input image is stored in the original image memory 52 (step 501), and the run extraction unit 53 extracts the run in the main scanning direction and stores it in the memory (step 502). The connected rectangle extracting unit 54 extracts a connected rectangle using a run on the memory, and the ID assigning unit 55 assigns a rectangle ID (serial number) to the connected rectangle, and the rectangle ID constitutes the connected rectangle component. All the runs are also given (step 503). The rectangle ID selection unit 56 selects a specific connected rectangle (rectangular ID) (that is, a processing target) (step 504), and the frequency counting unit 57 searches for a run having the specified rectangle ID, and the frequency. Are counted (step 505).

  Next, the threshold setting unit 58 obtains a threshold based on the distribution of run frequencies (step 506). The connected rectangle extraction unit 59 extracts connected rectangles for only the runs that are larger than the calculated threshold with respect to the extracted runs in the main scanning direction (step 507). The ruled line extraction unit 60 extracts a ruled line from the extracted connected rectangle (step 508). Similar processing is performed in the sub-scanning direction (step 510), and ruled lines are extracted.

  If there are three or more ruled lines in both the main scanning direction and the sub-scanning direction (step 511), the character rectangle determining unit 61 determines that the character rectangle is a character rectangle (step 512). The following processing is the same as in the second embodiment.

(Example 6)
In general, the run frequency distribution of the connected rectangular components of the table frame including the vertical and horizontal lines is as shown in FIG. That is, run lengths 1 to 10 are run distributions of vertical lines, and 10 to 28 are run distributions of characters in contact with vertical lines or horizontal lines. A run distribution of 29 or more is a horizontal run distribution. In the distribution of FIG. 14, by setting the threshold value to 29, only the horizontal line can be extracted. When the differential value of the distribution is zero, that is, when the run distribution does not change, that point becomes the threshold value. In this embodiment, the difference is used to search for this threshold value.

  FIG. 15 shows the configuration of the sixth embodiment. The difference from the fifth embodiment is that a difference calculation unit 65 is provided. FIG. 16 shows a process flowchart of the sixth embodiment.

  The difference calculation unit 65 obtains a difference from the adjacent frequency in order from the smallest run length in the frequency distribution (step 606). The threshold setting unit 58 sets the run length at which the difference is zero as the threshold (step 607). Thereafter, similarly to the fifth embodiment, the connected rectangle extracting unit 59 extracts connected rectangles only for runs larger than the set threshold in the main scanning direction (step 608), and the ruled line extracting unit 60 is extracted. Ruled lines are extracted from the connected rectangles (step 609).

(Example 7)
The run distribution of the document including the table created in the office has a tendency as shown in FIG. 14, but when the difference is obtained as in the above-described embodiment 6, even if the value is smaller than the run length value 29 due to noise or the like. May match the adjacent distribution frequency value. Alternatively, even a run length greater than 29 may be a frequency value of 10 or more as a frequency value, and the frequency value does not always coincide with the neighbor. This is due to high-frequency component noise in the run distribution.

  In general, high frequency component noise can be removed by a FIR (Finite Impulse Response) type digital fill. Therefore, in this embodiment, the digital filter is used to remove a portion corresponding to high frequency noise.

  FIG. 17 shows the configuration of the seventh embodiment, in which a filter processing unit 66 is further added to the configuration of the sixth embodiment. FIG. 18 shows a process flowchart of the seventh embodiment. Steps 701 to 705 and steps 707 to 712 are the same as the processing in the sixth embodiment. In step 706, the filter processing unit 66 applies a digital filter (low pass filter) to the frequency distribution to remove high frequency noise.

(Example 8)
FIG. 19 shows a run distribution with only horizontal lines. When the run distribution is taken in units of connected rectangles, a connected rectangle that forms a table frame and a connected rectangle that forms a horizontal line are included. In order to extract a connected rectangle of only horizontal lines in the vicinity of the threshold value 33, it is only necessary to find a point where the differential value becomes zero at a position larger than the peak of the run distribution.

  FIG. 20 shows the configuration of the eighth embodiment. The difference from the seventh embodiment is that the peak detector 67 is provided and the processing content of the difference calculator 65 is different. FIG. 21 is a process flowchart of the eighth embodiment.

  The processing up to step 806 is the same as in the seventh embodiment. In step 807, the peak detection unit 67 searches for a point where the secondary differential value is zero or the sign of the differential value changes from the one with the smaller run length in the frequency distribution, and sets it as a peak. Next, the difference calculation unit 65 obtains a difference from the adjacent frequency behind the peak (step 808). The threshold setting unit 58 sets the run length at which the difference is zero as the threshold (step 809). Since the following processing is the same as that of the seventh embodiment, the description thereof is omitted.

Example 9
When recognizing a table, it is necessary to repeatedly extract connected rectangles, and each time a run is extracted from an original picture and a connected rectangle is extracted, it takes time. Therefore, by preparing only the run information in advance, it is possible to shorten the processing time for extracting other feature amounts using the run.

  In other words, since the results of the process can be accumulated by holding the attributes of the run, the run that has been recognized can be excluded from the next recognition process. As a result, the processing time of the entire recognition process can be reduced. Shortening is possible. At the same time, since recognition is possible in units of runs, highly accurate recognition processing can be performed in every detail. Moreover, since it is converted into run information, various image processing can be performed in a short time.

  FIG. 22 shows a configuration of the ninth embodiment. In this embodiment, an attribute information recording unit 68 and a character data erasing unit 69 are further added to the configuration of the eighth embodiment. FIG. 23 is a process flowchart of the ninth embodiment. In step 903, the run extraction unit 53 secures an area for holding run attribute information (for example, attributes such as characters and lines) corresponding to the extracted run.

  The attribute information recording unit 68 records a mark indicating a character in a run constituting the connected rectangle in the rectangle determined as the character rectangle by the character rectangle determining unit 61 (step 918). After the character size is output, the character data erasing unit 69 examines the extracted run and erases the black pixel on the original image corresponding to the run to which the mark that is the character is given (step 922).

  In addition, as the run attribute information, an attribute indicating whether the run belongs to an image such as a line, a photograph, noise, line noise, or background may be held.

(Example 10)
The tenth embodiment is an embodiment in which the present invention is realized by software. FIG. 24 illustrates a system configuration example of the tenth embodiment. A recording medium such as a CD-ROM records the character size estimation processing function or processing procedure of the present invention and installs it in the system. A document set on a scanner or the like is read, a character rectangle is extracted from the document image developed on the memory, the size of the extracted character rectangle is estimated, and the result is displayed and output on a display or the like.

The structure of Example 1 of this invention is shown. The processing flowchart of Example 1 of this invention is shown. An example of the vertical size frequency distribution of the character rectangle containing two character sizes is shown. The structure of Example 2 of this invention is shown. The processing flowchart of Example 2 of this invention is shown. The structure of Example 3 of this invention is shown. The processing flowchart of Example 3 of the present invention is shown. FIG. 8 is a process flowchart subsequent to FIG. The structure of Example 4 of this invention is shown. The processing flowchart of Example 4 of the present invention is shown. An example of a core wire rectangle is shown. The structure of Example 5 of this invention is shown. The processing flowchart of Example 5 of the present invention is shown. Shows the frequency distribution of runs including a general table. The structure of Example 6 of this invention is shown. 9 shows a processing flowchart of Embodiment 6 of the present invention. The structure of Example 7 of this invention is shown. 9 shows a processing flowchart of Embodiment 7 of the present invention. The run distribution with only horizontal lines is shown. The structure of Example 8 of this invention is shown. 10 shows a processing flowchart of an eighth embodiment of the present invention. The structure of Example 9 of this invention is shown. The processing flowchart of Example 9 of this invention is shown. The structure of Example 10 of this invention is shown.

DESCRIPTION OF SYMBOLS 1 Image input part 2 Original picture memory 3 Run extraction part 4 Concatenated rectangle extraction part 5 Frequency counting part 6 Peak detection part 7 Character size output part

Claims (3)

Extracting a run in the main scanning direction and the sub-scanning direction of the input image; obtaining a frequency distribution of the extracted run; setting a predetermined threshold based on the frequency distribution of the run; Extracting a connected rectangle using a run larger than the predetermined threshold; extracting a ruled line from the extracted connected rectangle; and a predetermined number or more of ruled lines in the main scanning direction and the sub-scanning direction in a predetermined area. A step of determining a character rectangle when present, a step of obtaining a frequency distribution of the vertical size of the rectangle determined to be the character rectangle, a step of detecting a peak of the frequency distribution of the vertical size, and the peak A character size estimation method comprising a step of outputting, as a character size, a vertical size having a frequency value equal to or less than a predetermined value when searching in a direction in which the vertical size is large . Means for extracting runs in the main scanning direction and sub-scanning direction of the input image; means for obtaining the frequency distribution of the extracted runs; means for setting a predetermined threshold based on the frequency distribution of the runs; Means for extracting a connected rectangle using a run larger than the predetermined threshold; means for extracting a ruled line from the extracted connected rectangle; and a predetermined number or more of ruled lines in the main scanning direction and the sub-scanning direction in the predetermined area. Means for determining a character rectangle when present, means for determining a vertical size frequency distribution of the rectangle determined to be the character rectangle, means for detecting a peak of the vertical size frequency distribution, and A character size estimation device comprising means for outputting, as a character size, a vertical size having a frequency value equal to or less than a predetermined value when searching in a direction in which the vertical size is large . A computer-readable recording medium recording a program for causing a computer to implement the character size estimation method according to claim 1 .

Images